|
(11) | EP 0 742 547 A2 |
| (12) | EUROPEAN PATENT APPLICATION |
|
|
|
|
|||||||||||||||||||
| (54) | Method and apparatus for pattern recognition |
| (57) The method and apparatus performs pattern recognition on given information such as
speech data or image data with a reduced amount of calculations of the degree of matching
associated with reference patterns. The method and apparatus provides a high speed
operation without an increase in the amount of calculations of the degree of matching
even when there are a great number of reference patterns. In this invention, the process
of calculating the degree of matching includes: a scalar quantization step in which
the value of each dimension of an input feature vector is scalar-quantized; a table
look-up step in which a table is subjected to a look-up operation on the basis of
the quantized value obtained in the scalar quantization step so as to obtain an output
value of a reference pattern function for each dimension; and a cumulative matching
degree calculation step in which the output value of the reference pattern function
obtained in the table loop-up step is accumulated for all dimensions, and the resultant
cumulative value is finally employed as the degree of matching between the reference
pattern and the input feature vector. |
[0001] The present invention relates to a method and apparatus for performing pattern recognition on an input pattern.
[0002] More specifically, the present invention relates to a method and apparatus for performing pattern recognition on a pattern obtained from input speech information so as to produce a sentence corresponding to the input speech information.
[0003] Furthermore, the present invention relates to a method and apparatus for achieving accurate pattern recognition with a reduced amount of computing operations.
[0004] It is known in the art in pattern recognition technique that reference patterns are prepared in advance and a reference pattern which shows the best matching with the input pattern is selected and employed as the recognition result. In general, an input pattern can be represented by a feature vector including a plurality of feature values. On the other hand, the reference patterns can be represented as feature vectors representative of recognition results or represented by a function of the feature vector.
[0005] When the reference patterns are given as feature vectors representative of recognition results, the degree of matching between an input pattern and a reference pattern is represented by the distance between the feature vector associated with the input pattern and the feature vector associated with the reference pattern. On the other hand, if the reference patterns are given by a function of the feature vector, the degree of the matching between an input pattern and a reference pattern is represented by a value of the reference function of a given feature vector associated with an input pattern.
[0006] In many cases, the reference pattern function is given in a form of a multidimensional probability density function. If a given feature vector and the probability density function for an ith reference pattern are denoted by x and Pi(·), respectively, then the degree of matching Yi between the input pattern and the ith reference pattern is given by:
[0007] A function such as a Gaussian distribution function is employed as the probability density function. In some cases, a mixed density function such as the weighted sum of a plurality of probability density functions is also employed. When a mixed density function is employed, the degree of matching is represented by:
where Pim(·) is the mth probability density function associated with the ith reference pattern, and wim is the weight of the mth probability density function associated with the ith reference pattern.
[0008] Furthermore, when there is no correlation among the dimensions elements of the feature vector given by equation (1), the following function may also be employed as the reference pattern function:
where xj is the feature value of the jth-dimension element of an input vector x, and Pij(·) is the probability density function corresponding to the jth-dimension element of the ith reference pattern.
[0009] In speech recognition, a hidden Markov model (HMM) is usually employed. In this case, the reference patterns correspond to individual HMM states, and the each HMM state represents the output probability corresponding to the input pattern.
[0010] In practical pattern recognition, Pi(·) merely represents the degree of matching between an input pattern and a reference pattern, and thus Pi(·) is not necessarily required to be a probability density function in a rigorous sense. Pi(·) can be regarded as the distance between an input vector and a reference pattern vector. Furthermore, a usual function other than distance functions may also be employed as Pi(·). Thus, in the following description, the term "reference pattern function" or "matching degree function" is used instead of the term "probability density function".
[0011] A problem in the conventional technique of pattern recognition described above is that equation (1) has to be calculated for all reference pattern functions. In particular, when there are a great number of reference patterns, it is required to perform a great amount of calculations to determine the degree of matching.
[0012] It is an object of one aspect of the present invention to solve the above problem. More, it is an object of another aspect of the invention to provide a method and apparatus for performing high-speed pattern recognition with a reduced amount of calculations of the degree of matching associated with all reference patterns. It is an object of a further aspect of the invention to provide a method and apparatus for pattern recognition by which high-speed pattern recognition can be performed without an increase in the amount of calculations even when there are a great number of reference patterns. It is an object of another aspect of the invention to provide a method and apparatus for pattern recognition by which accurate pattern recognition can be performed without a significant increase in the amount of calculations of the degree of matching.
[0013] One aspect of the present invention provides a pattern recognition method for performing pattern recognition on -the basis of calculation of the degree of matching between an input feature vector and a reference pattern, the pattern recognition method being characterized in that: in the matching degree calculation process, the degree of matching between the scalar-quantized feature vector and the reference pattern is accumulated for all dimensions, and the resultant cumulative value is employed as the degree of matching between the reference pattern and the input feature vector. The pattern recognition method further comprises a matching degree pre-calculation step in which the degree of matching is pre-calculated for each dimension of the reference pattern on the basis of the quantized value of each dimension elements of the scalar-quantized feature vector, and the obtained result is stored, wherein in the matching degree calculation process, the stored degree of matching corresponding to the quantized value is read and accumulated.
[0014] In the pattern recognition method of one aspect of the invention, the matching degree pre-calculation step comprises: an approximate value calculation step in which an approximate value of the degree of matching between the quantized value of each dimension element and the value of the corresponding dimension of the reference pattern is calculated; an error calculation step in which a plurality of feature vectors are input, a calculation is performed so as to obtain an error of the degree of matching of the reference pattern calculated from each feature vector according to the above-described matching degree calculation process relative to the degree of matching of the reference pattern calculated in a more rigorous manner from the each feature vector, and the error is accumulated for all the plurality of feature vectors; and an optimization step in which the approximate value of the degree of matching is optimized so that the above-described error is minimized, and the optimized value is stored. Furthermore, in the above-described matching degree calculation process, the degree of matching of a reference pattern which is selected on the basis of the degree of matching calculated from a quantized value of each dimension element of a scalar-quantized feature value is replaced by a degree of matching which is calculated directly from the input feature vector without scalar quantization, and the replaced value is finally employed as the degree of matching between the reference pattern and the input feature vector.
[0015] According to another aspect of the present invention, there is provided a pattern recognition method for performing pattern recognition on the basis of calculation of the degree of matching between an input feature vector and a reference pattern, the pattern recognition method being characterized in that: the matching degree calculation process comprises: (a) a scalar quantization step in which the value of each dimension element of an input feature vector is scalar-quantized; (b) a table look-up step in which a table is subjected to a look-up operation on the basis of the quantized value obtained in the scalar quantization step so as to obtain an output value of a reference pattern function for each dimension; and (c) a cumulative matching degree calculation step in which the output value of the reference pattern function obtained in the table loop-up step is accumulated for all dimensions, and the resultant cumulative value is employed as the degree of matching between the reference pattern and the input feature vector.
[0016] In this pattern recognition method, the above-described table stores a pre-calculated value of the reference pattern function corresponding to the quantized value. Furthermore, the value stored in the above-described table is optimized in such a manner as to minimize the error of the degree of matching between the reference pattern calculated from each feature vector according to the matching degree calculation process relative to the degree of matching of the reference pattern calculated in a more rigorous manner from the each feature vector. Furthermore, in the scalar quantization step, the binary search method is employed. Furthermore, the above-described matching degree calculation process further comprises: (d) a matching degree re-calculation step in which a reference pattern is selected on the basis of the degree of matching calculated in the cumulative matching degree calculation step, and the degree of matching of the selected reference pattern is re-calculated from the input feature vector; and (e) a matching degree replacement step in which the degree of matching of the selected reference pattern is replaced by the degree of matching re-calculated in the matching degree re-calculation step. Furthermore, the above-described feature vector is a set of a plurality of feature vectors, and the matching degree calculation process calculates a set of degrees of matching, in the matching degree re-calculation step (d), a set of degrees of matching showing good matching as a whole is selected, and the degrees of matching are re-calculated from the input feature vector; and in the matching degree replacement step (E), the set of selected degrees of matching is replaced by the set of degrees of matching calculated in the matching degree re-calculation step. Furthermore, the reference pattern function is selected from probability density functions, wherein the probability density functions may include a Gaussian distribution function, and the Gaussian distribution function may be a mixed Gaussian density distribution function consisting of the weighted sum of a plurality of Gaussian distribution functions. The mixed Gaussian density distribution function is approximated by performing calculation with the assumption that all dimensions of the feature vector exhibit a mixed distribution independent of each other.
[0017] According to still another aspect of the invention, there is provided a pattern recognition apparatus for performing pattern recognition on the basis of calculation of the degree of matching between an input feature vector and a reference pattern, the pattern recognition apparatus comprising: scalar quantization means for scalar-quantizing the value of each dimension of an input feature vector and outputting the resultant quantized value; storage means for storing a pre-calculated output value of a reference pattern function corresponding to the quantized value; and matching degree calculation means for accumulating the output value corresponding to the quantized value for all dimensions of the reference pattern function obtained from the storage means and employing the resultant cumulative value as the degree of matching between the reference pattern and the input feature vector.
[0018] In this pattern recognition apparatus, the value stored in the storage means is optimized in such a manner as to minimize the error of the degree of matching of the reference pattern calculated from the scalar-quantized feature vector by the matching degree calculation means relative to the degree of matching of the reference pattern calculated in a more rigorous manner from the feature vector. The above-described scalar quantization means performs scalar quantization using the binary search method. The pattern recognition apparatus further comprises matching degree re-calculation means for re-calculating the degree of matching in such a manner that a reference pattern is selected on the basis of the calculated degree of matching, the degree of matching of the selected reference pattern is re-calculated directly from the input feature vector without performing scalar quantization, and the output value of the selected reference pattern is replaced by the re-calculated value. Furthermore, the above-described feature vector is a set of a plurality of feature vectors, and the matching degree calculation means calculates a set of degrees of matching. Furthermore, the reference pattern function is selected from probability density functions including Gaussian distribution functions and mixed Gaussian density distribution functions consisting of the weighted sum of a plurality of Gaussian distribution functions.
[0019] Embodiments of the present invention will now be described with reference to the accompanying drawings, in which:
Figure 1 is a block diagram illustrating the hardware construction of an embodiment of a pattern recognition apparatus according to the invention;
Figure 2 is a flow chart illustrating a procedure of pattern recognition according to the embodiment of the invention; and
Figure 3 is a schematic representation of an example of a binary searching process.
[0020] Figure 1 is a schematic diagram illustrating the hardware construction of an embodiment of a pattern recognition apparatus according to the present invention.
[0021] In Figure 1, reference numeral 1 denotes an output device such as a display or a printer (a laser beam printer or an ink-jet printer) for outputting a result of pattern recognition or a response corresponding to the result. Reference numeral 2 denotes an input device for inputting information such as speech or image information to be recognized wherein the input information is stored as an input pattern in a storage device 4. The speech information is input via a microphone while the image information is input via a scanner. Reference numeral 3 denotes a central processing unit (CPU) for performing numerical calculations and for controlling operations according to a procedure of the present embodiment stored in the form of a program in the storage device 4. The storage device 4 includes an external storage device such as a disk device and an internal memory device such as a RAM/ROM for storing: various variables and temporality values in the middle of calculation; a feature vector 4a associated with an input pattern; a set of code values 4b for scalar quantization; a reference pattern function 4c; a set of output values of the reference pattern function 4d; and a pattern recognition program 4e including a control program for performing the process shown in the flow chart of Figure 2. Data is transmitted between output device 1, input device 2, CPU3 and storage device 4 via computer bus 5.
[0022] In the present embodiment, the control program used by CPU3 to perform various processes may be pre-stored in storage device 4 included in the present apparatus or the control program may be such a program provided on an external storage device which can be mounted in a removable fashion on the apparatus of the invention. Otherwise, the control program may be such a program which is down-loaded from another apparatus via communication means (not shown) including a public communication line or an LAN.
[0024] Figure 2 is a flow chart illustrating the procedure according to the present embodiment of the invention. According to the procedure shown in this flow chart, An input pattern is input via the input device 2, and a feature vector is extracted from the input pattern. The data associated with the extracted feature vector is stored in a feature vector storage area 4a, which is subjected to a pattern recognition process according to the flow chart shown in Figure 2. In this embodiment, it is assumed that the number of the reference patterns is S and the dimension of the feature vector space is N. The reference patterns used in the process shown in the flow chart of Figure 2 are stored in a reference pattern function storage area 4c of the storage device 4. The number S of the reference patterns used in the calculation of the degree of matching with an input pattern may be equal to the total number of the reference patterns stored in the reference pattern function storage area 4c, or otherwise a part of the reference patterns may be employed. Thus, S is a variable number. Furthermore, the input feature vector is denoted by x, and its jth-dimension element is denoted by xj. The function corresponding to the ith reference pattern is represented by Yi = Pi(·). In this embodiment, the dimensions of the feature vector space are assumed to be independent of each other, and Pi(x) is given by:
or
where Pij(·) is the jth-dimension function of the ith reference pattern function. The difference between equations (4) and (5) is that the latter is given in a logarithmic form. In the following description, it is assumed that Pi(x) is given in the logarithmic form.
[0025] After performing initialization in steps S1 to S3, the jth-dimension value xj of the input vector x is scalar-quantized in a scalar quantization step S4. In this scalar quantization, xKjj having a value which is closest to the input value xj is selected from the set of Kj code values {x1j, x2j,..., xKjj} which have been prepared in advance and which are stored in the scalar quantization code value storage area 4b.
[0026] Then in a table look-up step S5, a table look-up operation is performed to determine the jth-dimension output value Pij(xKjj) of the reference pattern function corresponding to xKjj obtained in the scalar-quantization described above. In this process, since the set of possible input values {x1j, x2j,..., xKjj} is known, the set of possible output values {Pij(x1j), Pij(x2j),..., Pij(xKjj)} is prepared in advance in the form of a table. Thus, the output value can be obtained by looking up Kjth element of the table.
[0027] Next in a cumulative matching degree calculation step S6, the result obtained in the table look-up step S5 is added to the cumulative value yi which has been obtained in the previous operation at the table look-up step S5 in the loop operation.
[0028] The steps S4 to S6 are performed repeatedly N times so as to obtain an approximate value yi of the degree of matching Yi between the input vector and the ith reference pattern. The above N-iterative operations are controlled in steps S3, S7, and S8.
[0029] Thus, the degree of matching of the input vector can be obtained quickly for all reference patterns by repeating the above-described steps S2 to S10 S times. The steps S1, S9, and S10 are for controlling the S-iterative operations.
[0030] In the scalar quantization step S4, the binary search technique may be employed to achieve a high-speed operation in the scalar quantization. In the binary search technique, the elements of the set of code values are arranged in ascending or descending order, and it is first determined whether a given jth-dimension value is included in the first half or the second half of the set. If it turned out that the jth-dimension value is in the first half of the set, the first half of the set is taken as a new set of code values, and the value is searched for in a recursive manner while the second half of the set is taken as a new set of code values if it turned out that the jth-dimension value is in the second half of the set, as shown in Figure 3.
[0031] In Figure 3, a set of code values and an input value to be scalar-quantized are shown in block 301. Block 302 illustrates a procedure of the binary searching.
[0032] First, it is determined whether the input value "5" is included in the first half subset {1, 3, 4, 7} of the full set of code values {1, 3, 4, 7, 11, 12, 15, 19} or in the second half subset {11, 12, 15, 19} (step 1). Then it is further determined whether the input value is included in a subset {1, 3} or {4, 7} of the subset {1, 3, 4, 7}. Finally, it is determined which value of the subset {4, 7} is more proper as the scalar-quantized value (step 3). Thus, "4" is finally output as the quantized value.
[0033] Alternatively, after performing a high-speed calculation of the degree of matching yi according to the technique described above, some of yi may be replaced by a value obtained by performing a rigorous calculation of the degree of matching using the equation Yi = Pi(x) instead of employing a scalar-quantized value. In this case, which of yi should be re-calculated may be determined according to the value of yi.
[0034] For example, a predetermined number of yi which have shown the highest degrees of matching in the high-speed calculation may be taken for the re-calculation.
[0035] Alternatively, the re-calculation may also be performed as follows. In some cases, pattern recognition is performed by employing not simply a single feature vector but a set of a plurality of feature vectors {x0,..., xT}. In such cases, sets of degrees of matching {Yi0,..., YiT} are used to obtain a result of pattern recognition. In this pattern recognition technique, a set of degrees of matching {Yi0,..., yiT} showing a good result as a whole may be taken and re-calculated so as to obtain {Yi0,..., YiT}, while the other sets employ the values obtained by the high-speed calculations.
[0036] For example, in speech recognition, when the reference patterns include {a, o, oo, g, k, n, s, t, y}, if [t oo k y oo] is taken as a time series of combined reference patterns, first, a set of matching degrees {yi0,..., yiT} is obtained according to the high-speed calculation technique of the present embodiment. Even if the result shows a very high matching degree in [t], if the result shows a very low matching degree in [oo k y oo], it would be meaningless to perform a re-calculation on [t]. On the other hand, even if [t] shows a very low matching degree, if [oo k y oo] shows good matching, it is meaningful to perform a rigorous re-calculation on [t].
[0037] If re-calculation is performed on a selected reference pattern in the above-described manner to obtain a more rigorous degree of matching, it is possible to reduce the error due to the scalar quantization and also the error due to approximation based on for example equation (15) (in this equation, the calculation which cannot be separated into individual dimensions in a rigorous sense is performed separately for each dimension thereby obtaining an approximate result).
Example 1
[0038] In this specific example, a Gaussian distribution (multidimensional Gaussian distribution) function is employed as the reference pattern function.
[0039] The degree of matching Yi between an input x and an ith reference pattern is represented by a Gaussian distribution function N(·) having an average of µi and a covariance matrix Σi as described below.
[0041] In the above equation, N denotes the number of dimensions of feature vectors, and t is used to represent transposition.
[0042] Yi is calculated at a high speed using the technique of the above-described embodiment according to the invention.
[0043] With the assumption that the dimensions of the N-dimensional feature vector space are assumed to be independent, the equation (6) can be decomposed as described below:
where µij denotes the average value associated with the jth dimension of the Gaussian distribution regarding the ith reference pattern, and σij denotes the variance associated with the jth dimension of the Gaussian distribution regarding the ith reference pattern. Although the dispersion is usually denoted by σ2, it is denoted simply by σ in this invention.
[0044] Furthermore, a set of code values {xj.k} is prepared for use as values to which the jth-dimension value xj of an input vector x is scalar-quantized. For all values of i, j, and k,
is calculated and the result is stored in the form of a table in the storage device 4.
[0045] After the preprocessing described above, the degree of matching is calculated for pattern recognition as described below.
[0046] First, in step S4, a given input feature vector x is scalar-quantized for each dimension element. That is, an optimum value of Kj is determined for each dimension j, and then
[0047] Then in a table look-up step S5, yij.Kj corresponding to xj.Kj is searched for by means of a table look-up operation.
[0048] Furthermore, in a cumulative matching degree calculation step S6, the result obtained in the table look-up step S5 is added to the cumulative value thereby obtaining an approximate value of Yi as described by the following equation:
[0049] Thus, the degree of matching Yi between the input feature vector x and each reference pattern is obtained by means of the high-speed calculation technique of the invention.
[0050] As described above, yi which has been obtained for some reference patterns {i} by means of the high-speed calculation may be replaced by a value obtained by performing rigorous calculation of the degree of matching using the equation (6) or (7).
Example 2
[0051] In this example, a mixed Gaussian density distribution function is employed as the reference pattern function. The mixed Gaussian density distribution function refers to a weighted sum of M Gaussian distribution functions, and can be defined by:
where wm denotes a weighting factor for the mth Gaussian distribution function. If the non-diagonal covariance elements of the Gaussian distribution functions is equal to 0, then
[0052] The degree of matching is calculated according to the above mixed Gaussian density function using the technique of the above-described embodiment of the invention.
[0053] A first technique is to apply the calculation technique of the above-described embodiment to each Gaussian distribution function constituting the mixed Gaussian density distribution function.
[0054] First, a value xj.Kj of each dimension of an input vector x is obtained as in Embodiment 1 described above. Then,
is determined by means of a table look-up operation.
[0055] Furthermore, in a cumulative matching degree calculation step S6, yi is obtained by calculating a cumulative value according to the following equation:
[0058] First, in a scalar quantization step S4, scalar quantization is performed to obtain a value xj.Kj for each dimension of an input vector x. Then
is determined by means of a table look-up operation.
[0059] Furthermore, in a cumulative matching degree calculation step S6, yi is obtained by calculating a cumulative value according to the following equation:
[0060] A third technique is to perform learning using learning data so as to obtain a function to be used to output Pi(x).
[0061] First, a set of values {yij.k} corresponding to a set of scalar-quantized code values {xj.k} is prepared for all values of i via learning according to a rule described later.
[0062] First, in a scalar quantization step S4, scalar quantization is performed to obtain a value xj.Kj for each dimension of an input vector x. Then the degree of matching is calculated according to the following equation:
[0063] In the above calculation, {yij.k} corresponding to each value of i and {xj.k} is obtained using learning data as follows.
[0064] First, the square error (∈2)in of a vector ξn = (x1.Kn1,..., xN.KnN) obtained by scalar-quantizing an nth learning data ξn=(ξn1,..., ξnn) for each dimension relative to (y1.Kn1, ..., yj.KnN) corresponding to ξn is defined as follows:
[0065] This square error indicates an error between an approximate value of the degree of matching and a true value, which occurs when an approximate value of the degree of matching between the nth learning data and the ith reference pattern is calculated using a set of {yij.k}.
[0066] If the set of {yij.k} is determined so that the square error (∈2)in is minimized for all learning data which have been prepared in advance, the set of {yij.k} can be employed to calculate a good approximate value of the degree of matching. That is, the set of {yij.k} is determined so that
is minimized for n learning data. This can be achieved using a known technique for minimizing the error.
[0067] Alternatively, in equation (19), the square error (∈2)in may also be defined using an input vector ξn which is not scalar-quantized, as follows:
[0068] As described above, the set of {yij.k} can be constructed again by means of learning so that the error due to the approximation based on for example equation (15) as well as the error due to the scalar quantization can be reduced.
Example 3
[0069] In Example 1 or 2 described above, the value of the probability density function is directly employed. Alternatively, in this example, the logarithm of the probability density function is employed in the calculations.
[0070] If the logarithmic values are used, the multidimensional probability density function can be represented by the sum, instead of the product, of probability density functions of individual dimensions. For example, equation (7) can be rewritten as:
1. A pattern recognition method for performing pattern recognition, the pattern recognition
method comprising the steps of :
calculating the degree of matching between an input feature vector and a reference pattern wherein in the calculating step a degree of matching is performed between a scalar quantized feature vector and when a degree of matching is performed the reference pattern is accumulated in all dimensions; and
employing the resultant accumulated value as the degree of matching between a reference pattern and the input feature vector.
2. A pattern recognition method according to Claim 1, further comprising the steps of
degree-pre-calculation matching in which the degree of matching is pre-calculation
for each dimension of the reference pattern on the basis of the quantized value of
each dimension element of the scalar quantized feature vector; and
storing the obtained result wherein, in the matching degree calculation process, a stored degree of matching corresponding to the quantized value is read and accumulated.
storing the obtained result wherein, in the matching degree calculation process, a stored degree of matching corresponding to the quantized value is read and accumulated.
3. A pattern recognition method according to Claim 2, wherein in the step of degree-pre-calculation
matching comprises the steps of approximate-value calculating in which an approximate
value of the degree of matching between the quantized value of each dimension element
and the value of the corresponding dimension of the reference pattern is calculated;
error calculating for calculating plurality of feature vectors are input, the error of the degree of matching of the reference pattern calculated from each feature vector according to the matching degree calculation process relative to the degree of matching of the reference pattern calculated in a more rigorous manner from each of said feature vectors;
accumulating the error for all of the plurality fo feature vecors; and
performing an optimization step in which the approximate value of the degree of matching is optimized so that the error is minimized and the optimized value is stored.
4. A pattern recognition method according to Claim 1, wherein in the degree-calculation
matching step, the degree of matching of a reference pattern which is selected on
the basis of the degree of matching calculated from a quantized value of each dimension
element of a scalar quantized feature value is replaced by a degree of matching which
is calculated directly from the input feature vector without scalar quantization and
the replaced value is employed as the degree of matching between a reference pattern
and the input feature vector.
5. A pattern recognition method according to any preceding claim, wherein said pattern
is a pattern representing a feature of speech data.
6. A pattern recognition method according to any preceding claim, wherein said pattern
is a pattern representing a feature of image data.
7. A pattern recognition method according to any preceding claim, wherein the recognition
result obtained on the basis of the degree of matching between the reference pattern
and said input feature vector is displayed on display means.
8. A pattern recognition method according to any one of claims 1 to 6 wherein the recognition
result obtained on the basis of the degree of matching between the reference pattern
and said input feature vector is output to a printer.
9. A pattern recognition method according to Claim 8, wherein said printer is a laser
beam printer.
10. A pattern recognition method according to Claim 8, wherein said printer is an ink-jet
printer.
11. A pattern recognition apparatus for performing pattern recognition on the basis of
calculation of the degree of matching between an input feature vector and a reference
pattern, said pattern recognition apparatus comprising:
scalar quantization means for scalar-quantizing the value of each dimension of an input feature vector and outputting the resultant quantized value;
storage means for storing a pre-calculated output value of a reference pattern function corresponding to said quantized value; and
matching degree calculation means for accumulating the output value corresponding to said quantized value for all dimensions of the reference pattern function obtained from said storage means and employing said resultant cumulative value as the degree of matching between the reference pattern and the input feature vector.
12. A pattern recognition apparatus according to Claim 11, wherein the value stored in
said storage means is optimized in such a manner as to minimize the error of the degree
of matching of the reference pattern calculated from the scalar-quantized feature
vector by said matching degree calculation means relative to the degree of matching
of the reference pattern calculated in a more rigorous manner from the feature vector.
13. A pattern recognition apparatus according to Claim 11, wherein said scalar quantization
means performs scalar quantization using the binary search method.
14. A pattern recognition apparatus according to Claim 11, further comprising matching
degree re-calculation means for re-calculating the degree of matching in such a manner
that a reference pattern is selected on the basis of the calculated degree of matching,
the degree of matching of the selected reference pattern is re-calculated directly
from the input feature vector without performing scalar quantization, and the output
value of said selected reference pattern is replaced by said re-calculated value.
15. A pattern recognition apparatus according to Claim 11, wherein said feature vector
is a set of a plurality of feature vectors, and said matching degree calculation means
calculates a set of degrees of matching.
16. A pattern recognition apparatus according to Claim 11, wherein said reference pattern
function is selected from probability density functions including Gaussian distribution
functions and mixed Gaussian density distribution functions consisting of the weighted
sum of a plurality of Gaussian distribution functions.
17. A pattern recognition apparatus according to any one of claims 11 to 16, wherein said
feature vector to be subjected to pattern recognition is extracted from speech data.
18. A pattern recognition apparatus according to Claim 17, further comprising a microphone
for inputting said speech data.
19. A pattern recognition apparatus according to any one of claims 11 to 16, wherein said
feature vector to be subjected to pattern recognition is extracted from image data.
20. A pattern recognition apparatus according to Claim 19, further comprising a scanner
for inputting said image data.
21. A pattern recognition apparatus according to any one of claims 11 to 20, wherein a
recognition result of the input feature vector is determined on the basis of the degree
of matching between the reference pattern and the input feature vector calculated
by said matching degree calculation means, and said recognition result is displayed
on display means.
22. A pattern recognition apparatus according to any one of claims 11 to 20, wherein a
recognition result of the input feature vector is determined on the basis of the degree
of matching between the reference pattern and the input feature vector calculated
by said matching degree calculation means, and said recognition result is output to
a printer.
23. A pattern recognition apparatus according to Claim 22, wherein said printer is a laser
beam printer.
24. A pattern recognition apparatus according to Claim 22, wherein said printer is an
ink-jet printer.
25. A pattern recognition method according to any one of claims 1 to 10 including the
step of inputting data comprising a pattern having said input feature vector.
26. A pattern recognition apparatus according to any one of claims 11 to 24 including
input means for inputting data comprising a pattern having said input feature vector.
27. A storage device for storing instructions for carrying out the method of any one of
claims 1 to 10.
28. A storage device according to claim 27 comprising a machine readable storage medium
containing machine readable code.